Research map / Graph learning / Graph algorithms and clustering

EverybodyDance: Bipartite Graph–Based Identity Correspondence for Multi-Character Animation

NeurIPS 2025 · Graph learning › Graph algorithms and clustering · approach: Optimization algorithms for graph matching · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes EverybodyDance, a diffusion-based multi-character pose-driven animation method that enforces identity correspondence via a bipartite Identity Matching Graph with Mask-Query Attention edge weights, plus identity-embedded guidance, multi-scale matching, and pre-classified sampling.

Problem

Pose-driven character animation fails to maintain correct one-to-one identity correspondence between reference and generated characters in multi-character scenes, especially when characters swap positions or occlude each other.

Approach

Formalize identity correspondence as a graph structural ratio over a weighted complete bipartite graph between reference and generated character nodes (masked latent regions), with edge weights from mask-restricted cross-attention (Mask-Query Attention), and optimize it as a training loss alongside identity-colored skeleton guidance.

Key results

• On ICE-bench, EverybodyDance reaches FVD 225.06 vs 364.85 for AnimateAnyone (about 38.3% reduction) and SSIM 0.654 vs 0.623 for the best baseline UniAnimate.
• On the Follow-Your-Pose-V2 public multi-character benchmark, FVD 127.36 vs 142.76 for Follow-Your-Pose-V2.
• Ablation: removing IMG supervision (End2End-M) raises FVD to 285.09 from 225.06 in the full model; removing pre-classified sampling raises FVD to 252.33.

Measured effects reported in the paper

Surprising findings

• Fine-tuning AnimateAnyone on the authors' dataset without the proposed components slightly worsens several metrics versus the original AnimateAnyone (e.g., FVD 358.31 vs 364.85 is better, but SSIM 0.596 vs 0.616 is worse).

Limitations

• Cannot effectively handle severe inter-character occlusion.
• Performance depends on the accuracy of the upstream instance segmentation model.
• Quantitative evaluation relies on proxy video-fidelity metrics rather than dedicated IC metrics.

Experimental setup

Datasets: custom multi-character dataset (~800 video clips) Follow-Your-Pose-V2 public multi-character benchmark TikTok

Benchmarks: ICE-bench Follow-Your-Pose-V2 multi-character benchmark TikTok multi-character 3-5 person videos

Metrics: SSIM PSNR* LPIPS L1 FID FID-VID FVD

Baselines: MagicAnimate AnimateAnyone MagicPose MimicMotion Follow-Your-Pose-V2 UniAnimate DisCo MagicDance

Builds on: AnimateAnyone ReferenceNet DWPose Stable Diffusion

Closest papers by meaning

• Animate Anyone: Consistent and Controllable Image-to-Video Synthesis for Character Animation (CVPR 2024 · similarity 0.87)
• Animate-X: Universal Character Image Animation with Enhanced Motion Representation (ICLR 2025 · similarity 0.86)
• ReMask-Animate: Refined Character Image Animation Using Mask-Guided Adapters (AAAI 2025 · similarity 0.86)
• Towards Multiple Character Image Animation Through Enhancing Implicit Decoupling (ICLR 2025 · similarity 0.85)
• FaceShot: Bring Any Character into Life (ICLR 2025 · similarity 0.85)
• Free-viewpoint Human Animation with Pose-correlated Reference Selection (CVPR 2025 · similarity 0.85)